Smartphone chatting and payment habits help AI identify daily stress, small study finds

by The Korea Advanced Institute of Science and Technology (KAIST)

edited by Sadie Harley, reviewed by Andrew Zinin

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources Credit: Unsplash/CC0 Public Domain Clues to stress can be hidden in everyday habits such as chatting, searching and making payments on a smartphone. KAIST researchers have developed a technology that analyzes several routine behaviors repeatedly performed within apps to identify a person's daily stress state.

A research team led by Professor Uichin Lee from the Department of AI Computing developed "SHIRBT-MMF," a technology that automatically identifies repeated smartphone behaviors and uses factors such as the duration and frequency of each behavior and text-input patterns to determine daily stress status.

The research was published in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies.

Previous smartphone-based stress research has mainly analyzed factors such as screen time, the number of app launches, time and frequency of use by app category, and typing behavior. However, when the apps being used or their interfaces change, the data being analyzed also change, which can reduce the stability of models designed to track stress over long periods.

The research team therefore focused on "what a person does" rather than "which app they use." Even if a user switches from KakaoTalk to LINE or WhatsApp, the activity itself remains "messenger chatting." In other words, the analysis is based on the purpose of the behavior rather than the name of the app.

The team defined these repeated behaviors as "Smartphone Human Interaction Routine Behavior Tasks (SHIRBTs)." Using its independently developed "Multi-Level Sequential Pattern Mining (ML-SPM)" technique, the team identified recurring screen-transition patterns across multiple days and then used a large language model (LLM) to assign easily understandable behavior labels such as "messenger chatting," "web browsing," "financial payment" and "reading/writing email."

In evaluations of whether the same behavior could be consistently classified across different email, web browser and messenger apps, the system achieved an accuracy of more than 95%.

To protect privacy, the system did not store the actual text or message content entered by users, recording only behavioral information such as whether text was entered, how often it was entered, and screen transitions. Information about message recipients or call partners was also processed using hashing, which converts the original information into values that do not directly reveal it.

The research team analyzed real-life data from 26 university students and office workers in their 20s and 30s. Among participants for whom both self-reported stress assessments and smartphone records were available, 16 short-term participants provided an average of 30.5 days of data, while 10 long-term participants provided an average of 105.1 days.

Based on the participants' daily self-assessments, the researchers built personalized models that distinguished between days with and without stress.

A total of 29 types of repeated behaviors were identified in participants' smartphone use, with each person routinely repeating an average of about seven behaviors. The research team used personalized artificial intelligence models trained on these behavioral patterns to determine whether a person was experiencing stress on a given day.

The best-performing model achieved an average accuracy of 75.9% for short-term participants and 73.8% for long-term participants, meaning it was correct roughly three out of four times. These results were 5.5 percentage points and 3.5 percentage points higher, respectively, than the conventional approach of analyzing app usage using the same AI technique.

Performance could also be maintained even when the number of behaviors used in the analysis was reduced. With the same personalized model, using only two behaviors, "messenger chatting" and "financial payment," or even analyzing "messenger chatting" alone, produced higher average classification accuracy than the conventional method.

In the long-term data, analytical indicators based on repeated behaviors also changed less over time and maintained a more stable structure than those used in the conventional approach.

In this experiment, the researchers collected a broad range of smartphone usage records to compare their method with conventional approaches. In the future, the team expects that once an individual's repeated behaviors have been identified, focusing data collection and analysis on those relevant behaviors could reduce the burden of data collection.

The study suggests the potential for developing digital mental health monitoring technology that makes it easier to understand which behaviors were used in classification and remains relatively stable even when the app environment changes.

Professor Lee of the KAIST AI Computing Department said, "We viewed the smartphone not simply as a device for measuring usage time, but as a space in everyday life where an individual's communication, information seeking, and financial activities are repeatedly carried out," adding, "This study provides a foundation for converting complex smartphone records into human-understandable behavioral units and offering long-term mental health tracking more reliably."

However, this study used associations between behavior and stress and did not demonstrate a causal relationship showing that a particular behavior causes or reduces stress. Based on findings from Android smartphone users in their 20s and 30s, the research team plans to expand the study to a broader range of age groups, occupations and smartphone environments and verify its long-term stability.

Hansoo Lee was the study's first author, professor Uichin Lee was the corresponding author, and researchers Taehyeon Park and Youngji Koh, along with professor Jae-Gil Lee of the School of Computing, were co-authors.

Hansoo Lee et al., Leveraging Smartphone Human Interaction Routine Behavior Task Mining and Modeling for Daily Stress Monitoring, Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (2025). DOI: 10.1145/3770644 Key concepts AI alignment Human-centered AI interfaces Digital privacy risks Agentic consumer AI AI surveillance capitalism Provided by The Korea Advanced Institute of Science and Technology (KAIST) Who's behind this story? Sadie Harley BSc Life Sciences & Ecology. Microbiology lab background with pharmaceutical news experience in oil, gas, and renewable industries. Full profile →

Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →